Fig. 1

Expression of underlying task knowledge is context-dependent. a Behavioral schematic in the reinforced context. b Mice are trained to lick to the target tone for water and to withhold from licking to the foil tone. c Behavioral sensitivity (d’) over trials in the reinforced context (max d’: 2.7 ± 0.2, N = 14 mice, black line is sigmoidal fit to average). d Hit and false-alarm rates of individual animals at peak performance rates (hit rate: 96.0 ± 0.9%, N = 14 mice; false-alarm rate: 28.0 ± 2.9%). e Top: same as a. Middle: 20–40 probe trials are interleaved with reinforced training each day, during which the licktube was removed and no reinforcements available. Bottom: typical daily training structure, 200–300 reinforced trials, 20–40 probe trials, 70–200 reinforced trials. f Still-frames from a movie in a training session between trials 1500–2000. Reinforced context (licktube present); mouse correctly responding to a target tone; mouse erroneously responding to a foil tone. Probe context (no licktube present, same session): mouse correctly responding to a target tone in the probe context with a lick; mouse correctly withholding a response to a foil tone. g Top: average hit rate (95.8 ± 1.4%, N = 14 mice) and false-alarm rate (81.6 ± 4.6%) across trials 1500–2000 in the reinforced context. Bottom: average hit rate (92.8 ± 2.1%, N = 14 mice) and false-alarm rate (18.7 ± 3.5%) in the probe context. F(3,39) = 198.05, one-way repeated-measures ANOVA, Tukey’s post-hoc correction, p = 0.84 between hit rates, p < 0.05 for all other comparisons. h Hit rates remain constant across contexts (Trials 1500–2000, hit rates, pre-reinforced = 91.3 ± 3.21%, probe = 89.1 ± 2.92%, post-reinforced = 89.7 ± 2.90%, n = 14 animals; F(2,26) = 0.1932, p = 0.851 between pre-reinforced and probe contexts, p = 0.914 between reinforced contexts, p = 0.973 between post-reinforced and probe contexts, one-way repeated measures ANOVA, Tukey’s post-hoc correction). i False-alarm rate during probe trials is significantly lower than during the reinforced trials (Trials 1500–2000, pre-reinforced false-alarm rate: 81.9 ± 4.5%, probe false-alarm rate: 15.5 ± 2.7%, post-reinforced false-alarm rate: 77.7 ± 4.9%, n = 14 animals; F(2,26) = 107.8, p < 0.0001 between reinforced sessions and probe contexts, p = 0.548 between reinforced contexts, one-way repeated measures ANOVA, Tukey’s post-hoc correction). j Behavioral sensitivity (d’) is significantly higher during probe trials than in the reinforced sessions (Trials 1500–2000, pre-reinforced d’: 0.40 ± 0.18, probe d’: 2.46 ± 0.18, post-reinforced d’: 0.53 ± 0.19, N = 14 mice; F(2,26) = 44.56, p < 0.001 between reinforced sessions and probe contexts, p = 0.827 between reinforced contexts, one-way repeated measures ANOVA, Tukey’s post-hoc correction). k Learning trajectories of an individual animal in the reinforced (black, n = 24 training sessions) and probe (gray, n = 6 training sessions) context. Dots indicate individual training sessions; lines indicate a sigmoidal fit. l Average d’ of a subset of animals whose learning was tracked in both the reinforced (black, N = 4–7 mice per time bin) and probe (gray, N = 4–7 mice per time bin) contexts. Dots indicate trial bins; solid lines indicate a sigmoidal fit. Reinforced trials to expert: 4728 ± 647 trials; probe trials to expert: 1765 ± 108 trials; N = 7 mice, t(6) = 4.359, p = 0.0055. m Average learning trajectories in the reinforced context (left) and probe context (right) (N = 7 mice, magenta = hit rate, cyan = false-alarm rate, all error bars indicate mean ± s.e.m)